Artificial T1-Weighted Postcontrast Brain MRI: A Deep Learning Method for Contrast Signal Extraction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Reducing gadolinium-based contrast agents is crucial for cost, environmental, and patient safety reasons.
• Published methods for synthesizing contrast-enhanced images have not been comparatively evaluated.

Purpose of the Study:

• To compare deep learning methods for synthesizing artificial T1-weighted full-dose MRI images from noncontrast and low-dose images.
• To evaluate a proposed contrast signal extraction method against existing state-of-the-art techniques.

Main Methods:

• A prospective study involving 213 participants undergoing brain MRI with low-dose (0.02 mmol/kg) and full-dose (0.1 mmol/kg) contrast.
• Two deep learning methods (Settings A & B) and a proposed method (Setting C) were reimplemented.
• Artificial and true full-dose images were compared by two readers assessing lesion interchangeability, enhancement, and conformity.

Main Results:

• The proposed method (Setting C) demonstrated significantly higher interchangeability (70/100 scans) compared to Settings A (40/100) and B (57/100).
• Setting C produced the smallest mean enhancement reduction in lesions (-0.50 ± 0.55) versus true images.
• Conformity scores for lesions were highest with Setting C (2.48 ± 0.91) compared to A (1.75 ± 1.07) and B (2.19 ± 1.04).

Conclusions:

• The proposed contrast signal extraction method significantly improves the synthesis of postcontrast MRI images.
• Despite improvements, a notable proportion of synthesized images still show inadequate interchangeability with reference true-dose images.